Title: Harnessing and Leveraging Stability in Latent Flow Matching

Abstract:

This study demonstrates that Latent Flow-Matching (LFM) models exhibit significant resilience to various perturbations, such as reductions in data volume and decreases in model capacity. We define this stability through the models' propensity to produce consistent results when subjected to identical noise seeds. By connecting this observation to flow matching theory, we argue that such stability is an intrinsic characteristic of the FM objective. Building on this insight, we develop practical methodologies to enhance both training efficiency and inference speed.

Specifically, our first finding reveals that LFM models maintain high performance even when trained on substantially smaller datasets. In scenarios with limited computational resources, these models achieve faster convergence without compromising output quality. This approach offers several benefits, including reduced training durations and diminished annotation requirements for conditional models.

Secondly, the stability of LFM under architectural reduction enables a novel two-stage, coarse-to-fine inference strategy. This method employs a lightweight architecture for the initial phase of the flow matching trajectory and switches to a higher-capacity model for the subsequent phase, significantly lowering inference costs. To identify the most valuable samples for this process, we propose three sample-scoring metrics and assess their effectiveness using standard generative model benchmarks. Our comprehensive evaluation across multiple datasets confirms the practical utility of this stability, highlighting advantages such as reduced data dependency and an inference speed improvement of more than two times, all while preserving output quality.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC